Research

On the same GPU, what wall-time cost is required to reach a given numerical error with Ramanujan CUDA versus simulated Quantum Amplitude Estimation?

Compare Ramanujan's 1914 pi series in a CUDA C++ kernel with Quantum Amplitude Estimation in CUDA-Q/cuQuantum, validated against exact SymPy ground truth.

On an RTX 5070 Laptop GPU, the CUDA implementation reached double-precision saturation (~16 digits) in ~2.6 ms, while simulated QAE reached ~5 digits in ~0.44 s. No crossover was observed on this hardware.

Run the optional H100 axis, then open the hardware-agnostic benchmark format to community submissions.

How do GGUF quantization and CPU/GPU layer placement change inference throughput and memory use on consumer NVIDIA hardware?

Separate prefill and decode with llama.cpp counters; discard warmups; report repeated-run variance; verify memory clock state; reject unstable or paged runs.

Across three 7B model families, decode throughput fell monotonically as quantized weight size increased. Full Llama-2 offload measured 6.2x CPU-only decode throughput.

Add batch-size and context-length sweeps, quality regression by quantization format, and a datacenter-GPU comparison.

Where do GPU implementations outperform CPU baselines across common algorithms and input sizes?

Run configurable CPU/GPU benchmark sweeps, export the measurements to CSV, and inspect empirical complexity and speedup in an interactive dashboard.

In the published sweep, radix sort is the strongest GPU result, while BFS and reductions remain slower than their CPU baselines at the tested sizes.

Extend the documented benchmark protocol to additional algorithms, input shapes, and verified GPU runs.